美国收紧AI芯片出口管制,限制中国企业远程使用海外算力
US tightens AI chip export controls to deter Chinese companies from accessing compute remotely
- 1美国商务部出台临时新规,数据中心若可能向中国等受限国家企业提供算力,出口芯片和服务器需申请许可
- 2新规覆盖超1200万TPP的设施,约合750块Nvidia H100,间接访问与中介渠道同样受限
- 3马来西亚、新加坡、印尼、泰国和巴西被新增许可要求,美方称这些国家存在较高芯片转售风险
- 4该规则还废除拜登政府的AI Diffusion Rule,现有运营商须在生效后90天内取得许可
管控从限制芯片流向延伸到限制算力使用方式,远程调用海外数据中心算力也被纳入许可体系,绕道获取算力的路径被收紧。
- 影响:依赖境外算力训练模型的中国企业,获取海外数据中心服务的合规成本和不确定性上升;第三国数据中心与出口商需核查客户的股权结构和实际用途;在东南亚等地部署算力的许可流程也会变长。
- 参考:可考虑先盘点当前租用的境外算力规模、供应商所在地和合同条款,评估是否落入许可范围;在东南亚或巴西布局算力的企业,值得提前测算许可申请的时间成本;数据中心运营方可梳理D:5背景客户的股权清单。
- 依据:规则内容以商务部临时规则文本为准;黄仁勋辞职与Megaspeed或被列入实体清单均为媒体报道口径,暂无独立验证。
The Trump admin on Monday is preparing to roll out new export controls on AI chips and data centers outside the U.S., seeking to prevent companies, especially those in China, from getting past U.S. export controls by remotely accessing AI infrastructure located in other countries.
Under the interim rule, the Commerce Department will require companies to get licenses to export, reexport or transfer AI chips and servers to data centers anywhere in the world if they may provide compute to companies in China or other restricted countries.
The remote-access rules target companies based in the U.S.’s D:5 group of embargoed countries, which includes China, Russia, Iran, North Korea and 18 others.
The U.S. already has rules requiring chip companies to get a license to ship AI chips to companies based in, or owned by any business in, this group of countries.
“This new control is a logical extension of existing controls,” the interim rule reads. “These restrictions will undermine the ability of companies headquartered in countries of concern from using U.S. advanced Super Intelligence (SI) semiconductors to train advanced SI models.”
The move comes days after Nvidia co-founder Jensen Huang resigned amid a scandal involving an alleged Chinese spy linked to one of the chipmaker’s customers. The U.S. government has reportedly been wanting to put that customer, Megaspeed, on a special entity list, as it believes the company has been buying thousands of Nvidia chips to provide compute to Chinese businesses to get past existing export-control restrictions.
Reuters reported last week that Chinese AI companies are rushing to secure computing capacity overseas amid fears of tighter restrictions from the U.S.
The rule applies to data centers or any network that can supply over 12 million total processing performance (TPP), an export-control metric the U.S. uses to measure computing power. (Twelve million TPP roughly works out to the max computing power of about 750 Nvidia H100 or H200 chips). The rule also covers indirect access through go-betweens or brokers, and exporters and data center operators are required to keep an eye on their customers’ ownership structures and activities to ensure they aren’t run by anyone from the D:5 group.
The rule also adds new licensing requirements for chip exports, reexports or transfer of advanced semiconductors to Malaysia, Singapore, Indonesia, Thailand and Brazil. The Commerce Department said it had included these as they “present a higher-than-normal risk of diversion.” For these countries, however, the license requirements will not apply to chips “destined” for companies based in the U.S.
Existing data center operators providing services that fall under this new rule are required to get licenses within 90 days of the rule coming into effect.
Notably, the rule doesn’t apply to commercial recommendation systems, like those used in e-commerce or social media.
The rule also rescinds the Biden admin’s AI Diffusion Rule from January 2025, which sought to impose country-level limits on exports of advanced AI chips, and restrict transfers of certain powerful AI models. That May, the Trump admin said it would throw out that framework, arguing that it added unnecessary restrictions on U.S. companies and the country’s allies.
“By making these changes, this IFR prevents the use of certain SI items in data centers and at a computing power level that can be used for training or running of advanced SI models,” the rule reads. “Access to such computing power can be used to train and run advanced SI models capable of supporting military- and intelligence-related activities that could threaten U.S. national security or foreign policy interests”
When you purchase through links in our articles, we may earn a small commission. This doesn’t affect our editorial independence.
Ram is a financial and tech reporter and editor. He covered North American and European M&A, equity, regulatory news and debt markets at Reuters and Acuris Global, and has also written about travel, tourism, entertainment and books.
You can contact or verify outreach from Ram by emailing [email protected].
来源:TechCrunch · AI · techcrunch.com